Abishek Giri

Backend & distributed systems · Wichita, KS

I'm a computer-science graduate student at Wichita State. I build backend systems, storage, consensus, ledgers, schedulers, and I care most about how they behave when things go wrong: retries, races, partial failure, and whether you can still prove the system is correct afterward. Most of these I wrote from scratch to actually understand them, not from tutorials.

Open to Summer 2027 software-engineering internships.

Selected work

• ShardDB

  Go · Raft · 2PC

  A distributed, strongly consistent key-value store written from scratch in Go. It implements Raft properly, leader election, log replication, and snapshotting with a real InstallSnapshot RPC, plus two-phase commit for multi-key transactions and per-key locking. I verified it Jepsen-style: killing leaders, partitioning the network, and checking a sum invariant holds under concurrent load and crash recovery.

  Source →

• MIDAS Ledger

  Python · FastAPI · PostgreSQL

  A concurrency-safe double-entry ledger. Transfers are atomic and idempotent by key; concurrency is handled with SELECT ... FOR UPDATE and deterministic lock ordering so balances can't drift under parallel load. Under extreme contention it returns 503 rather than risk a partial write, explicit failure over silent corruption. No UI, on purpose; the point is the guarantees.

  Source →

• Job Orchestrator

  Python · PostgreSQL

  A multi-tenant job orchestrator with race-free leasing (FOR UPDATE SKIP LOCKED), a transactional outbox for exactly-once delivery, a reaper that recovers failed leases, and HMAC-signed worker polling. It ships with scripts that actually prove the properties, e.g. no job is double-claimed across N workers.

  Source →

• AgentLens

  Python · developer tool

  Local trace capture and evidence-based diagnosis for AI-agent failures. When a run breaks it names the specific step, shows the evidence behind the diagnosis, and suggests a fix, or reports insufficient evidence instead of guessing. Two lines instrument a run; the demo reproduces a broken agent locally without calling a provider. Early (v0.1.x).

  Source →

• PermissionDiff

  Python · on PyPI

  A CI-native tool that shows how a code change alters effective authorization, did this pull request quietly grant a user, tenant, or service access it didn't have before? It generates authorization cases from declared subjects/resources/actions and checks invariants; verdicts are deterministic Python, never LLM judgments, and failing cases are recorded for reproduction.

  Source → PyPI →

• ReleaseGate

  Python · release governance

  Evidence infrastructure for software releases: deterministic, versioned policy decisions with signed, offline-verifiable proof of what was allowed and why, plus a customer-installable GitHub Action that signs and verifies deploys against a tenant.

  Source →

Open source

Merged fixes and documentation to iii-hq/iii, plus pull requests to OpenSearch, Kubernetes, FastAPI, Dapr, Bazel, GraphQL.js, and Appwrite, mostly input validation, correctness, and runtime behavior. Full activity →

Research & experiments

• Large-Scale Bug Prediction. An end-to-end research project analyzing more than 700,000 source files across 10 major Python repositories (Django, NumPy, Pandas, Flask) for bug prediction.
• AI in Software Engineering. Research on applying AI techniques to software-engineering problems.
• Research Paper. An academic research paper repository.
• VLM Failure Modes. An investigation of whether Vision-Language Model uncertainty (such as entropy) actually rises on noisy, ambiguous, or corrupted inputs in safety-critical settings.
• VLM Adversarial Defense. An investigation of the adversarial robustness of Vision-Language Models (LLaVA-1.
• Amnesic VLM Defense. An investigation of why adversarial hallucinations occur in VLMs, using entropy analysis, attention visualization, and focal ablation to show they are confident, robust, and distributed rather than localized.
• Sentiment Steering GPT. Activation steering on GPT-2 Medium to identify and manipulate a moral direction, inverting outputs on negative concepts without retraining.

Engineering investigations, not peer-reviewed publications.

Other projects

• Adaptive Release Controller Python · CI/CD
• Change Risk Predictor Python · Machine Learning
• Travel Buddies React · Node.js
• StorePro Fullstack React · Node.js
• NYC Taxi Pipeline Streamlit · Python
• HighPerf Parquet C++ · C
• LLM-CODEGEN LLM · Generative AI
• Data Watermarking Engine Python · Security
• PlanetStore Dashboard React · Dashboard
• Distributed KV Store Distributed Systems · Go
• Real-Time Chat WebSockets · Socket.IO
• AI Recommender Recommender Systems · Python
• Pathfinding Visualizer Algorithms · Visualization
• Dementia Data Processor Healthcare · Data Analysis
• Python Search Engine Information Retrieval · Algorithms
• BasketMind Sports Analytics · Data Science
• Navigation System Graph Theory · Routing
• WeatherBuddy API Integration · Frontend
• Bayes Gaussian Plot Bayesian · Math
• Statistics Calculator Statistics · Math
• Portfolio Website Flask · Tailwind CSS
• s3 Storage · AWS
• AI Impact on Jobs Data Analysis · Research
• Streaming Analytics Platform Flink · Kafka
• Distributed KV Raft Distributed Systems · Raft
• Adaptive Layout Engine CSS · Layout
• Criteo CTR Pipeline ML · CTR
• PermissionDiff Python · Security
• MLSys 2026 Scheduler Python · Scheduling

Experience & education

• Aug 2026 - Present

  M.S. Computer Science

  Wichita State University
• Sep 2022 - May 2026

  B.S. Computer Science

  Stockton University
• 2023, 2024, 2025

  Summer Entrepreneurship and Business Academy (SEBA)

  Stockton University
• 2026

  Open Source Contributor

  GitHub · Python, Go, TypeScript, Java
• Aug 2023 - May 2026

  Resident Assistant

  Stockton University
• Sep 2022 - May 2025

  Math Tutor

  Stockton University

Contact

Best way to reach me is email: giriabishekkumar5@gmail.com. I'm also on GitHub and LinkedIn.